Sunteți pe pagina 1din 6

2011 9th IEEE International Conference on MonB4.

3
Control and Automation (ICCA)
Santiago, Chile, December 19-21, 2011

A Systematic Tunning Method of PID Controller


for Robot Manipulators
Xiaoou Li, Wen Yu

Abstract This paper addresses the iterative tuning Most robot manipulators employed in industrial oper-
method of PID control for the robot manipulator based on ations are controlled by PID algorithms independently at
the responses of the closed loop system. Several properties each joint [23]. Compared with the above linear systems-
of the robot control are used, such as any PD control can
stabilize a robot in regulation case, the colsed-loop system based tuning algorithms, there are some difficulties to
of PID control can be approximated by a linear system, design a systematic tuning method for robot PID control
and the control torque to the robot manipulator is linearly Robot manipulators are strong nonlinear systems, and
independent of the robot dynamic. By using these properties,
a novel systematic turning method for the PID control is the torque of one joint affects the other and vice versa.
proposed. Simulations and experimental results of an upper If the gains are tuned heuristically [25], Cohen-Coon
limb exoskeleton give validation of this PID tuning method. method [4] and optimization [12] methods. There are
too many gains to tune simultaneously for robot. A
I. I NTRODUCTION 6-degrees-of-freedom robot manipulator has 18 gains
to be tuned. When one gain is tuned, it requires to
The proportinal-integral-derivative (PID) control has tune the other 17 gains in turn because of dynamics
simple structure and clear physical meanings for its three coupling in robot.
gains. The control performances are acceptable in the most Based on stability analysis, the upper bounds of
of industrial processes. It has been used in more than PD gains and lower bound of derivative gain can
90% of various practical control systems [1][2]. Three be derived. However, these bounds cannot guarantee
parameters of PID controller are tuned such that the desired performances.
performances at transient, including rise-time, overshoot,
and settling time, steady-state error, are satisfied, mean- There are few research regarding PID gains tuning
while the closed-loop system is stable and robust against for robot manipulators. PID tuning algorithms cannot be
plant modeling uncertainty and disturbances. The study on used straight because the responses are nonlinear. The
tuning methods of PID controller mainly focused on linear intelligent techniques have been applied for PID gains
systems [19]. The tuning methods for PID controllers can tuning, for example fuzzy logic [22], neural networks
be grouped according to their nature and usage: [14] and genetic method [11], but the final controllers are
Analytical methods: PID parameters are calculated
no longer linear PID, they become complete intelligent
from analytical or algebraic relations between a plant control systems. Another PID tuning method for robots
model and an objective [5][6][16]. is impedance control [13], which first uses inverse dy-
Heuristic methods: These are evolved from practical
namics to transfer the robot into a linear system. Then
experience in manual tuning [25][4][2], and from some mechanical impedance ideas are applied to tune
artificial intelligence techniques [22][14][11]. PID gains. In [5] discrete-time approximation of inverse
Frequency response methods: frequency characteris-
dynamics was calculated such that PID parameters could
tics of the controlled process are used to tune the PID be adjusted. Lyapunov approach was used in [7] to adjust
controller [20]. PID controller such that it follows linearization control. All
Optimization methods: These can be regarded as a
above methods need the models of robot manipulators, and
special type of optimal control, where PID parameters their PID controllers do not have clear physical meaning.
are obtained ad hoc using an offline numerical In this paper, three important properties of PID control
optimization [12]. of robot manipulator are applied for PID gains turning.
Adaptive tuning methods: These are for automated 1) Any PD controller can stabilize a robot in regulation
online tuning, using one or a combination of the case when its gains are positive
previous methods based on real-time identification 2) The behavior of the colsed-loop system of PID
[24]. control is simple, and it can be approximated by
a linear system
Xiaoou Li is with the Departamento de Computacion, CINVESTAV-
IPN, Mexico City, Mexico.Wen Yu is with the Departamento de Control 3) The control torque to the robot manipulator is inde-
Automatico, CINVESTAV-IPN, Mexico City, Mexico pendent of the other robot dynamic.
978-1-4577-1476-4/11/$26.00 2011 IEEE 274
MonB4.3

By using these properties, we propose a new systematic ball whose radius decreases approximately 1
, see
min ( )
tuning method for PID control. The turning steps are as Lewis et al. (2004). Theoretically, PD control is sufficient
follows for robot control. However, in order to decrease steady-
1) a) Stabilize the robot with a PD control state error caused by gravity and friction, derived gain
b) Add a step input to the closed-loop system in has to be increased. The closed-loop system become slow.
(a), and save the step response. Usually, the big settling time does not allow us to increase
c) Search a linear time-invariant model, which has as we want.
a similar step response with (b). Although adding an integrator can extraordinarily de-
d) Tune PD/PID gains similar with the linear crease steady-state error, the overshoot of the closed-loop
system in (c) system becomes larger and robustness property deterio-
e) Refine PID gains in (d) by prior knowledge. rates.
Finally, we apply this method on an upper limb ex-
oskeleton. The experimental results show this PID tuning A. Tuning in closed-loop
method is effective for robot manipulator Since it is danger to send a step command to the joints
of the exoskeleton robot. We use closed-loop identification
II. PID TUNING FOR ROBOT MANIPULATORS and tuning method. Here we use two properties of the
The dynamics of robots are derived from Euler- robot dynamics:
Lagrange equation. It can be written as The control torque of the robot is dependent of the
other terms;
() + ( ) + () + () = (1)
PID control is linear.
where represents the link positions. is joint num-

It is well known the robot (1) is open-loop unstable, and
ber, () is the inertia matrix, ( ) = { } positive gains of a PD controller can guarantee closed-
represents centrifugal force, () is vector of gravity loop stability (bounded) in regulation case , see Spong
torques, () is unknown disturbance. is control and Vidyasagar (1989). We first use a PD control (2) with
input. = 0 and small and , to stabilize the robot.
Classical linear PID law is When the desired position is constant, the closed-loop
Z system is stable,
= + ( ) + (2)
0 () + ( ) + () + () = 0
where = is desired joint angle, and

Considering gravity compensation, the closed-loop system
are proportional, integral and derivative gains of the is
PID controller, respectively. This PID control law can be
() + ( ) + () + () = 0 () (5)
expressed via the following equations
= + + where () = () + () () is estimated gravity.
(3) Now we will use a tuning rule to find another PID
= (0) = 0
controller 1 for this closed-loop system. If we define
It is known that in regulation case, any positive gains the final control torque as
of the PD controller
= 1 + 0 ()
= + (4)
Obviously, the closed-loop system is
can guarantee stability of the closed-loop system, see
() + ( ) + ()+() 0 + () = (6)
Spong and Vidyasagar (1989). PD control does not guar-
antee the achievement of the position control objective The control for the closed-loop system is
because manipulators dynamics contain the gravitational = 1
torques vector, unless gravity compensation is applied. The
integrator is the most effective tool to eliminate steady- Since the PID control is linear, this idea can be extended
state error, in this way PD control (4) becomes PID to general case,
control (2). However, integrator gain has to be increased X
when the robot is heavy. This causes big overshoot, long () + ( ) + () + () = ()
settling time, and less robust. An approximation model =1
compensation can decrease integrator gain, see Kelly et where is tuning times, and
al. (2005). X X X Z
X
It is known that if the PD control law (4) is applied to = + ( ) +
each joint, the position tracking error is bounded within a =1 =1 =1 0 =1

275
MonB4.3

This means we can start from small PID gains to stabilize robot is Taylor series model as in (9). The model can be
the robot first, then tuning the other PID controllers written in frequency domain
independently. The final PID control is the summarization
()
of all these controllers. = 2 2 (11)
() + 2 + 1
or
B. Linearization of the colsed-loop system ()
=
There are several methods to linearize robot models. If () (1 + 1 ) (1 + 2 )
the velocity and gravity are neglected, the terms ( )
The responses of this second order model are similar with
and () in the nonlinear dynamics (1) are zero, resulting
mechanical motions. If there exists a big overshoot, a
in a linear model of the form,m see Goldenberg and
negative zero is added in (11)
Bazerghi (1986)
() = (7) () (1 + 3 )
= (12)
() (1 + 1 ) (1 + 2 )
It is an oversimplified model and is impossible for PID
tuning, because velocity and gravity are main control The normal input signals for PID tuning are step and
issues of robots. Most of robot, the gravity loading is a repeat inputs.
dominant component of the dynamics.
The velocity dependent term ( ) representing C. PD/PID tuning
Coriolis-centrifugal forces, can be assumed to be negli-
gible for small joint velocities. This is a rate linearization The linear PID law in time domain (2) can be trans-
scheme, see Golla et al. (1981), which results in a linear formed into frequency domain

model of the form 1
() = 1 + + () = () ()
+ = (8)

where = () |=0 = ()
0 is operating where = is proportional gain, = is integral

|=0
point. But many experiments, see Swarup and Gopal time constant and = is derivative time constant.
(1993), showed that even at low speeds ( ) should Because the robot can be approximated by a linear
be accounted for. system. Some tuning rules for linear systems can be
When the robot model is completely known, Taylor applied for the colsed-loop system tuning. We first give
series expansion can be applied, see Li (1989). At the PD tuning rules. When each joint can be approximated by
operating point 0 the nonlinear model (1) can be approx- a first-order system,
imated by
+ + = (9) =
1 +
where = () |=0 = [()+(
)]
|=0 = The PD gains are tuned as in Table 1, here Model 1 is from
()
|=0 Huang et al. (2005), Model 2 is from Chien and Fruehauf
Although the physical and mathematical structure of (1990).
the complete dynamic robot model is analytically coupled Table 1. PD turning for the first-order model
and nonlinear, the observed transient response of robot
dynamics appears to resemble the transient response of
Ziegler-Nichols tuning 05
the linear systems. Consequently, each joint of the robot
4 4
can be characterized as a single input-single output (SISO) Cohen-Coon method
3 + 4 11 +2
system. In this paper, we use this identification-based Our Method
2
1

linearization method. For each joint, typical linear model
is a first order system with transportation delay as Here and are obtained from Figure 1.
In Table 1 we list Ziegler-Nichols and Cohen-Coon

= (10) methods, they are PI controllers. From the best of our
1 + knowledge, PD turning rules are still not published and
The response is characterized by three parameters, the applied. If each joint is approximated by a second-order
plant gain the delay time and the time constant system,
. These are found by drawing a tangent to the step ()
= 2 2
response at its point of inflection and noting its intersec- () + 2 + 1
tions with the time axis and the steady state value.
The PD gains are tuned as in Table 2.
Sometimes the first model cannot describe the complete
nonlinear dynamic of robot. A reasonable linear model of Table 2. PD tuning for the second-order model
276
MonB4.3

Capsule/yuw/PID/JointSpace/srf1.wmf
is to decrease integrator gain, such that overshoot
y (t )
is small
Step 2 PD control 0 : use small PD gain to generate
a stable closed-loop system.
Km Step 3 Obtain the step responses of the closed-loop
system for each joint. Now the robot has been
compensated by gravity model, i.e.
() + ( ) + () = 0 ()
Step 4 Use the first-order or the second-order linear
t models to approximate the step responses of the
Tm closed-loop systems. PID gains are obtained by
m Table 1-Table 4, it is 1
Step 5 Refine PID gains with Table , it is 2
Fig. 1. Step response of a linear system Step 6 The final control for the robot is
= 0 + 1 + 2 ()

Method 1
51 1 +01 III. A PPLICATION TO AN EXOSKELETON
3 081
Method 2
2
1 Exoskeletons are wearable robots, which are worn by

2 the human operators as orthotic devices. The exoskeleton
Our Method 1
links, joints and work space correspond to those of the
When PD control cannot provide good performances, human body. The system may be used as a human input
PID control should be used. The PID gains for the first- device for tele operation, human-amplifier, and physical
order model is decided by Table 3. therapy modality as part of the rehabilitation process [10].
Although great progress has been made in a century-long
Table 3. PID tuning for the first-order model
effort to design and implement robotic exoskeletons, many

design challenges continue to limit the performance of
Ziegler-Nichols tuning 2 05 the system. One of the limiting factors is the lack of
(32 +6 ) 4
Cohen-Coon method
4 13 +8 11 +2 simple and effective control systems for the exoskeleton.
Our Method
2
2 1 The PID/PD control is the simplest scheme that can be
used to control robot manipulators. The exoskeletons are
The PID gains for the second-order model is decided
usually heavy, it is not easy to obtain an ideal PID for
by Table 4.
an exoskeleton robot. The 7-DOF upper limb exoskeleton
Table 4. PID tuning for the second-order model shown in Figure 2 is composed of a 3-DOF shoulder
(J1-J3), a 1-DOF elbow (J4) and a 3-DOF wrist (J5-J7).
Method 1
51
2 1 1 +01 J1-J3 are responsible for shoulder flexion-extension, ab-
3 081
2 ductionadduction and internalexternal rotation, J4 create
Method 2 2 1

2

elbow flexion-extension, J5-J7 are responsible for wrist
20
Our Method
15 10 flexionextension, pronation-supination and radialulnar
If the above four tables cannot give us good perfor- deviation.
mances, we use Table 5 to refine PID gains as 2 . The computer control platform of the UCSC 7-DOF
exoskeleton robot is a PC104 with an Intel Pentium4@2.4
Table 5. Effects of PID gains GHz processor and 512 Mb RAM. The motors for the first
Rise Overshoot Settling Steady Error Stability
four joints are mounted in the base such that large mass
P Decrease Increase
Small
Decrease Degrade
of the motors can be removed. Torque transmission from
Increase the motors to the joints is achieved using a cable system.
I
Small
Increase Increase
Large
Degrade
The other three small motors are mounted in link five.
Decrease Decrease The real-time control program operated in Windows XP
D
Small
Decrease Decrease
Minor
Improve
with Matlab 7.1, Windows Real-Time Target and C++ .
Decrease Decrease All of the controllers employed a sampling frequency of
The procedure of PD/PID tuning for robot control is 1. The properties of the exoskeleton with respect to
described as follows base frame are shown in Table 8.
Step 1 Gravity modeling (): the objective of this step Table 8. Parameters of the exoskeleton
277
MonB4.3

Capsule/yuw/PID/Exoskeleton/exosf1.wmf Capsule/yuw/PID/Exoskeleton/autof4.wmf
A ngle (r ad)
A ng le (r ad)
0 .3
0 .2

0 .2

Joint 1

Joint 2
0 .1 0 .1

a7
0
Time ( sec o nd) T im e (se con d)
1. 7 1 .8 1 .9 2 0. 8 0 .9 1 1 .1

0. 3
a4 Ang le (r ad) A ngle (ra d)

d1 0 .3

0. 2
a3

Joint 4
d5

Joint 3
0 .2
d3 0. 1

0 .1

0
T im e (se cond ) T ime ( sec on d)
0 2 .1 2. 2 2. 3 1. 9 1 .9 5 2 2 .0 5

Fig. 3. PD control of the exoskeleton and step responses of linear


models

Fig. 2. The UCSC 7-DOF exoskeleton robot.

The gravity compensation in (5) is calculated by () =



()
Joint Mass (kg) Center (m) Length (m) We will design a PID tuning rule for these linear
1 3.4 .3 .7 systems and apply the tuned PID controllers to the robot.
2 1.7 .05 .1 In order to tuning PID gains for the linear systems (14),
3 .7 .1 0.2 we rewrite the PID (2) as
4 1.2 .02 .05 Z
1
5 1.8 .02 .05 = + ( ) +
0
6 .2 .04 .1
7 .5 .02 .05 where = is proportional gain, = is integral


The two theorems in this paper give sufficient conditions time constant and = is derivative time constant.
for the minimal values of proportional and derivative gains We use the following tuning rule
and maximal values of integral gains. We first use the 20 2

following PD control to stabilize the robot = = 15 = (15)
10
= [150 150 100 150 100 100 100]
(13) to tune the PID parameters. This rule is similar with Huang
= [330 330 300 320 320 300 300] et al. (2005), and Chien and Fruehauf (1990), in their case
= 5 1 1 +01
3 = 21 = 081 It is
The joint velocities are estimated by the standard filters
different with the other two famous rules, Ziegler-Nichols

e 18 and Cohen-Coon methods, where = =
() = () = ()

+ + 30 2 = 05 or = 4
+ =
3 4
The PD regulation of the first four joints are shown (32 +6 )
13 +8 = Because their rules are 4
11 +2
in the sold lines of Figure 3. Then we use open-loop step suitable for the process control, our rule is for mechanical
responses of linear systems to approximate the closed-loop systems.
responses of the robot. By the rule (15), the PID1 gains are
1 = 602093
+9+1 1 = 90 1 = 1 1 = 540
1 2 = 30 2 = 2 2 = 60
2 = 202 +3+1
09 (14) (16)
3 = 552 +4+1 3 = 40 3 = 20 3 = 20
4 = 302085 4 = 90 4 = 15 4 = 270
+8+1

The step responses of the following four linear system are We apply these PID controllers 1 to the robot, the
shown in the dash lines of Figure 3. new closed-loop system
Here the main weight of the exoskeleton is in the first () + ( ) + () 0 + () = 1
four joints. The potential energy is
The control torque becomes
= 1 1 1 + 2 (1 1 + 2 2 1 ) + 3 3 1 2
+4 [4 4 (1 3 + 2 3 1 ) + 4 4 (1 3 2 1 3 ) + 3 1 2 ] = 1 + 0 () (17)
278
MonB4.3

Capsule/yuw/PID/Exoskeleton/autof5.wmf
this method on an upper limb exoskeleton, real experiment
0 .3 A n g le (r ad ) results give validation of our PID tuning method.

T h eo ry g ai n s P ID 0 R EFERENCES
0 .2 A u to - tu n in g P ID t
[1] K.H.Ang, G.Chong, Y.Li, PID Control System Analysis, Design,
F i n e tu n in g P ID 1 and Technology, IEEE Trans. Control System Technology, vol. 13,
no. 4, pp. 559576, 2005.
0 .1 R ef er en ce [2] K.J. Astrom, T. Hagglund, Revisiting the ZieglerNichols step
response method for PID control, Journal of Process Control, 14,
635650, 2004
[3] P. Cominos, N. Munro, PID controllers: recent tuning methods and
0 design to specifications, IEE Proceedings-Control Theory and
T im e (s e co n d ) Applications, 149 (1), 4653, 2002
[4] G.H. Cohen, G.A. Coon, Theoretical consideration of retarded
1 .2 1 .3 1 .4 1. 5 1. 6
control, Trans. ASME, 75, 827834, 1953
[5] P.H.Chang, and J.H.Jung, A systematic method for gain selection of
robust PID control for nonlinear plants of second-order controller
canonical form, IEEE Trans. Control System Technology, vol. 17,
Fig. 4. PID tuning for joint 1 (J1) no. 2, pp. 473483, 2009.
[6] I.-L.Chien, P.S.Fruehauf, Consider IMC tuning to improve con-
troller performance, Chemical Engineering Progress, 33-41, 1990
Capsule/yuw/PID/Exoskeleton/autof6.wmf [7] W.-D.Chang, R.C.Hwang, J.G.Hsieh, A self-tuning PID control for
Angle (rad)
a class of nonlinear systems based on the Lyapunov approach,
Angle (rad)
0.3 0.3 Journal of Process Control, 12, 233242, 2002
Fine tuning [8] A.A.Goldenberg, A.Bazerghi, Synthesis of robot control for assem-
0.2
Fine tuning
0.2 bly processes, Mech. Machine Theory, 21(1), 43-62, 1986
Referen
ce [9] D.F.Golla, S.C.Garg, P.C.Hughes, Linear state-feedback control of
Referenc
e
0.1
manipulators, Mech.Machine Theory, 16, 93-103, 1981.
0.1
[10] H.Herr, Exoskeletons and orthoses: classification, design challenges
and future directions, Journal of NeuroEngineering and Rehabili-
0
0
Time (second) Time (second) tation, 6(21) 2009.
1 1.05 1.1 1.15 1.05 1.1 1.15 1.2 1.25 1.3 [11] J-G.Juang, M-T.Huang, W-K.Liu, PID control using presearched
Joint 2
Joint 4 genetic algorithms for a MIMO system, IEEE Trans. Syst., Man,
Cybern. C, Appl. Rev., vol. 38, no.5, 716-727, 2008 .
[12] B. Kristiansson, B. Lennartsson, Robust and optimal tuning of
Fig. 5. Final PID control. PI and PID controllers, IEE Proceedings-Control Theory and
Applications, 149 (1), 1725, 2002
[13] N.Hogan, Impedance control: An approach to manipulation, PartsI
III, ASME J. Dynam. Syst., Meas., Contr., vol. 107, pp. 124, 1985.
Since we use linear PID control, the gains of and [14] F.L.Lewis, K.Liu, and A.Yesildirek, Neural net robot controller with
0 can be added together. The control results of Joint guaranteed tracking performance, IEEE Trans. Neural Networks,
vol. 6, no. 3, pp. 703715, 1995
1 are shown in Figure 4. [15] F.L.Lewis, D.M.Dawson, C.T.Abdallah, Robot Manipulator Con-
After this refine turning, the final PID control is 2 trol: Theory and Practice, Marcel Dekker, Inc, New York, NY
10016, 2004
their gains are [16] H.-P.Huang, J.-C.Jeng, K.-Y.Luo, Auto-tune system using single-
run relay feedback test and model-based controller design, Journal
= [5 4 5 6 3 4 2] of Process Control, 15, 713-727, 2005
= [320 280 210 250 210 210 220] [17] R.Kelly, V.Santibez, L.Perez, Control of Robot Manipulators in
= [410 400 420 430 410 410 410] Joint Space, Springer-Verlag London, 2005.
[18] C.J.Li, An efficient method for linearization of dynamic models of
The final control is robot manipulators, IEEE Transactions on Robotics and Automa-
tion, Volume: 5 Issue: 4 , 397 - 408, 1989
= 0 + 1 + 2 () [19] A.ODwyer, Handbook of PI and PID Controller Tuning Rules,
Imperial College Press, London, 2006
[20] F.G.Shinskey, Process Control Systems - Application, Design and
The control results are shown in Figure 4 and Figure 5. Tuning, McGraw-Hill Inc., New York, 1996
[21] A.Swarup, M.Gopal, Comparative study on linearized robot mod-
IV. C ONCLUSIONS els, Journal of Intelligent and Robotic Systems, Volume 7, Number
3, 287-297, 1993
In this paper, a new systematic tuning method for PID [22] Y. L. Sun and M. J. Er, Hybrid fuzzy control of robotics systems,
IEEE Trans. Fuzzy Systems, vol. 12, no. 6, pp. 755765, Dec. 2004.
control is proposed. This method can be applied to any [23] M.W.Spong and M.Vidyasagar, Robot Dynamics and Control, John
robot manipulator. By using several properties of robot Wiley & Sons Inc., Canada, 1989.
manipulators, the tuning process becomes simple and is [24] Q.-G.Wang, , Y.Zhang, X.Guo, Robust closed-loop identification
with application to auto-tuning, Journal of Process Control, 11,
easily applied in real applications. Some concepts for PID pp. 519-530, 2001
tuning are novel, such as step responses for the closed- [25] J. G. Ziegler and N. B. Nichols, Optimum settings for automatic
loop systems under any PD control, and the joint torque is controllers, Trans. ASME, vol. 64, pp. 759768, 1942
separated into several independent PIDs. We finally apply
279